Tomas et al.:
Tomas Hlasny wrote:
Thank`s to everybody who already provided help on this topic.
We used SAS PROC MIXED, as you recomended (but some older version) to
carry the ANOVA with autocorrelated data.
A caution: it is my recollection that PROC MIXED in some earlier
versions of SAS (what version did you use?) was troubled by some
bugs...When you wrote in question 1 that "only the latter one worked
properly," what did you mean?
Could somebody help me with these points?
1. What indicates which procedure on calculation of DDDF is
appropriate - the options are KW (Kenward and Roger), Sattertwaith
(modification of KW) and RESIDUALS.
Only the latter one worked properly - MODEL DDFM = Residuals. We used
it just by trial and error, do not understand the differences between
the methods. Can somebody briefly explain?
1. In my understanding (I am a soil scientist with a good grounding in
statistics, but not a statistician...), this is not a simple issue. In
Proc MIXED and linear mixed models in general, null distributions of the
test statistics are often unknown, and p-values cannot be computed
exactly. Kenward-Roger was recommended to us because it inflates
(appropriately) the estimated variance-covariance matrix of the fixed
and random effects. Descriptions of the various DDFM options in SAS PROC
MIXED are available in the SAS online documentation
(http://support.sas.com/onlinedoc/913/docMainpage.jsp). An interesting
discussion of these issues and a comparison of the various options can
be found in "Approximations to Distributions of Test Statistics in
Complex Mixed Linear Models Using SAS® Proc MIXED" (Schaalje et al.)
available at http://www2.sas.com/proceedings/sugi26/p262-26.pdf.
2. The result of the analysis is just like of normal ANOVA - there is
statistical difference between the groups, i.e. at least one differ
from the others (not very usefull info). Thus I need to carry out some
post hoc test (Duncan, Tukey ...) - is there some way how to do it in
the case of autocorrelated data?
2. In PROC MIXED, you can use LSMEANS statements with the appropriate
options (e.g., PDIFF, ADJUST) to calculate the least-square estimated
means for fixed effects in the model and request multiple comparisons
among/between them. CONTRAST and ESTIMATE statements can also be used to
generate appropriate comparisons among treatments or fixed effect factor
levels.
--
Jeffrey G. White, Ph.D.
Assistant Professor
Dept. of Soil Science
3207 Williams Hall
North Carolina State University
Campus Box 7619
Raleigh, NC 27695-7619
Tel: 919-515-2389 Fax: 919-515-2167
email: [EMAIL PROTECTED]
